Monocular Object SLAM using Quadrics and Landmark Reference Map for Outdoor UAV Applications

Localization is one of the core technologies commonly applied in Unmanned Aerial Vehicles (UAVs) in urban fields. Most UAVs use Global Navigation Satellite System (GNSS) to estimate their positions, whereas the GNSS signal can be affected by the occlusion of buildings in urban areas. We use a visual positioning method based on the monocular camera sensor information and offline reference map in the GNSS-denied environments to address this problem. Our goal is to obtain the UAV's pose data and the corresponding GPS position data by matching the simultaneous map with the reference landmark map. In order to adapt to the flight environment of different altitudes and different perspectives, we propose a monocular semantic SLAM system to provide long-range visual tracking and mapping of landmark buildings together with the feature points. Our SLAM system's landmark object is described in quadratic form, which contains information about its 3D position and poses to increase awareness of the environment. The system we propose provides positioning results with meter-level accuracy and improves the impact of accumulated visual SLAM errors.